Twitter | Pretraživanje | |
hardmaru
“AMD is aiming to disrupt 4K gaming, and build a Radeon GPU line-up to take on Nvidia in a similar way to how the firm has ramped up Ryzen processors to beat out Intel” What about OpenCL and deep learning?
Exec claims that Big Navi will disrupt 4K gaming, similarly to how Ryzen now dominates desktop CPUs
TechRadar TechRadar @techradar
Reply Retweet Označi sa "sviđa mi se" More
Jonathan Fly 👾 2. velj
Odgovor korisniku/ci @hardmaru @AMD
Every cool new project is tested and developed with NVIDIA, or sometimes TPUs. AMD would need to step up so much, form a team to port popular projects like GPT-2, StyleGAN, etc, really get out there. If they're doing it I'm not seeing it, while NVIDIA is constantly engaged.
Reply Retweet Označi sa "sviđa mi se"
lab member 001 2. velj
Odgovor korisniku/ci @jonathanfly @hardmaru @AMD
gpt-2 works fine on amd. stylegan2 does not
Reply Retweet Označi sa "sviđa mi se"
choongng 2. velj
Odgovor korisniku/ci @hardmaru @AMD
I’m pretty sure AMD is deliberately ceding deep learning and GPU compute to NVIDIA so they can focus on winning server CPU and graphics markets. They are much smaller than Intel+NVIDIA so I’m not sure they’re wrong but it sure is disappointing.
Reply Retweet Označi sa "sviđa mi se"
Timothy Liu 2. velj
Odgovor korisniku/ci @choongng @hardmaru @AMD
They need to deliver the software alongside ORNL Frontier (~2021). Effectively, US Govt has given them a significant cash boost to jump-start development. That said, not infallible (see original Aurora which failed to delivery in 2018)
Reply Retweet Označi sa "sviđa mi se"
Timothy Liu 2. velj
Odgovor korisniku/ci @hardmaru @AMD
AMD’s strategy seems to be Vega+ROCm to tackle compute workloads. So far, ROCm support hasn’t made it to any Navi cards, and it seems upcoming supercomputers are all Vega. And they’re not done building up ROCm software to match CUDA in features/stability/performance
Reply Retweet Označi sa "sviđa mi se"
Maxime Lenormand 2. velj
Odgovor korisniku/ci @hardmaru @AMD
As much as I'm all for competition and having multiple companies in one sector, I feel like Nvidia are taking such a huge step ahead with CUDA and now Rapids the hardware just isn't enough. But I think a lot of us would love higher vRAM consumer GPUs
Reply Retweet Označi sa "sviđa mi se"
Quentin Leon 🧢 2. velj
Odgovor korisniku/ci @hardmaru @AMD
Was wondering about this as well. AMD's Threadripper line of CPUs is massive for high-volume data pre-processing, and would love to pair that with an AMD GPU one day, but as it stands the performance and library support for DL tasks just aren't there yet.
Reply Retweet Označi sa "sviđa mi se"
Mutiny 2. velj
Odgovor korisniku/ci @hardmaru @AMD
Well unfortunately they barely on the map and I complained to AMD sales rep many times in the past but their ecosystem is way behind. If I would be the initial step I would do is similar to Intel did and is deploying some optimized inference backend that supports
Reply Retweet Označi sa "sviđa mi se"
Han 2. velj
Odgovor korisniku/ci @hardmaru @AMD
No MacOS support planned, which implies no cuDNN equivalent support. On the positive side, pricing pressured mainstream GPUs.
Reply Retweet Označi sa "sviđa mi se"
Igor Moura 2. velj
Odgovor korisniku/ci @hardmaru @AMD
When I got into ML and had to give away my AMD gpus in order to buy one from nvidia just to speed up the stuff I was learning made me really sad
Reply Retweet Označi sa "sviđa mi se"